Yeah, but this version is not necessarily optimal. If you have 4 positive samples, then you will test 2+2+1+1+1+1 = 8, instead of 4 in the naive case.

It feels like a problem for information theory.